JP2014106961A - Method executed by computer for automatically recognizing text in arabic, and computer program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To extract a text feature appropriately in recognizing a text in Arabic.SOLUTION: A two-dimensional array of pixels associated with pixel values each of which is expressed in a binary number is formed as a result of digitalization of a line of Arabic characters. The pixel values are expressed in a binary number. The line of Arabic characters is divided into a plurality of line images, and a plurality of cells are defined in one of the plurality of line images. Each of the plurality of cells has an adjoining pixel group. A two-value cell number is formed as a result of serialization of the pixel value of each pixel of the plurality of cells in one of the plurality of line images. A text feature vector is formed in accordance with the two-value cell number obtained from the plurality of cells in one of the plurality of line images. The text feature vector is sent to a hidden Markov model so that the line of Arabic characters is recognized.

Description

  This patent application was filed on Dec. 14, 2011 by the same inventor and is based on the same assignee's co-pending US patent application No. 13 / 325,789, entitled “Effective Extraction of Arabic Text Features”. , A system and method for recognizing Arabic text "and claims priority to the application. U.S. Patent Application No. 13 / 325,789 is filed by the same inventor on U.S. Patent Application No. 12 / 430,773, filed April 27, 2009, entitled "Effective Arabic Text Feature Extraction". Based on "System and Method for Arabic Text Recognition", the disclosure of which is hereby incorporated by reference.

BACKGROUND OF THE INVENTION This application relates generally to automatic recognition of Arabic text.

  Text recognition, i.e. automatic reading of text, is a field of pattern recognition. The purpose of text recognition is to read printed text with human accuracy and faster. Many text recognition methods assume that the text can be separated into individual characters. Such techniques work well for Latin words typed or typed, but cannot be reliably applied to cursive letters such as Arabic. Previous studies on the recognition of hand-drawn text in Arabic have found difficulties in attempting to segment Arabic words into individual characters.

  Different classification systems such as statistical models have been applied to the recognition of Arabic text. However, properly extracting text features remains a major obstacle in achieving accurate Arabic text recognition.

SUMMARY OF THE INVENTION In a general aspect, the present invention relates to a method for automatically recognizing Arabic text. The method obtains a text image that includes lines of Arabic characters and digitizes the lines of Arabic characters to form a two-dimensional array of pixels, each associated with a pixel value. The pixel values are expressed in binary numbers, and the method further includes dividing the line of the Arabic character into a plurality of line images and a plurality of lines in one of the plurality of line images. Each of the plurality of cells has a group of adjacent pixels, and the method further includes the step of: for each pixel of the plurality of cells in one of the plurality of line images. Form binary cell numbers by serializing pixel values and form text feature vectors according to binary cell numbers obtained from multiple cells in one of multiple line images Including a Rukoto, and recognizing the Arabic character line by sending a hidden Markov model and the text feature vectors (Hidden Markov Model).

  In another general aspect, the present invention relates to a method for automatically recognizing Arabic text. The method obtains a text image including a line of Arabic characters, digitizes the line of Arabic characters, thereby each of two-dimensional pixels associated with pixel values expressed in binary numbers The array of two-dimensional pixels includes a plurality of rows in a first direction and a plurality of columns in a second direction, and the method is further the same in the columns of pixels By counting the frequency of successive pixels with pixel values, forming a text feature vector using the frequency count obtained from the column of pixels, and sending the text feature vector to the hidden Markov model Recognizing the line of letters of the word.

  In another general aspect, the present invention relates to a method for automatically recognizing Arabic text. The method forms a two-dimensional array of pixels, each associated with a pixel value, by acquiring a text image that includes lines of Arabic characters and digitizing the lines of Arabic characters. And dividing the line of Arabic characters into a plurality of line images, generating a miniaturized line image by miniaturizing at least one of the plurality of line images, Forming a series of serialized numbers by serializing the pixel values of the pixels of each column of the digitized line image, the series of serialized numbers forming a text feature vector; The method further includes recognizing a line of Arabic characters by sending the text feature vector to a hidden Markov model.

  In another general aspect, the present invention relates to a computer program that includes a computer-readable program code function that causes a computer to obtain a text image that includes a line of Arabic characters, By digitizing the character lines, a two-dimensional array of pixels, each associated with a pixel value, is formed, the pixel values are represented in binary numbers, and the code function is further transmitted to the computer, further to the Arabic language. Dividing the word character line into a plurality of line images, defining a plurality of cells in one of the plurality of line images, each of the plurality of cells having a group of adjacent images, the code function The computer further stores the pixel value of each pixel of the plurality of cells in one of the plurality of line images. Arabic by forming a text feature vector according to binary cell numbers obtained from a plurality of cells in one of a plurality of line images and sending the text feature vector to a hidden Markov model Recognize lines of characters.

  System implementations may include one or more of those shown below. The method further converts a binary cell number to a decimal cell number and serializes the decimal cell numbers obtained from multiple cells in one of the multiple line images. Forming a series of decimal cell numbers, and forming a text feature vector according to the series of decimal cell numbers obtained from the plurality of cells in one of the line images. obtain. The two-dimensional array of pixels may include a plurality of rows in the first direction and a plurality of columns in the second direction. The line of Arabic characters can be aligned substantially along the first direction. The plurality of line images may be continuously arranged along the first direction. At least one of the plurality of line images may have a height defined by M rows in the first direction and a width defined by N columns in the second direction. M and N are integers. The two-dimensional array of pixels may include N rows of pixels. N may be in the range between 2 and approximately 100. N may be in the range between 3 and approximately 10. Pixel values in a two-dimensional array of pixels can be represented by a single bit binary number. The pixel values in the two-dimensional pixel array can be expressed in multi-bit binary numbers. The hidden Markov model can be implemented as a hidden Markov model toolkit.

  The systems and methods described in this application provide a comprehensive, quantitative and accurate technique for feature extraction in Arabic text. The disclosed Arabic character recognition is more efficient and requires less computation time than some prior art. The disclosed systems and methods are also simpler and easier to use than some prior art.

  While the invention has been particularly shown and described by reference to a plurality of embodiments, it will be understood by those skilled in the art that various changes and details in form can be made without departing from the spirit and scope of the invention. Will be understood.

BRIEF DESCRIPTION OF THE DRAWINGS The following drawings are incorporated into and form a part of the application documents, illustrate embodiments of the invention, and together with the description, serve to explain the nature of the invention.

It is a flowchart for demonstrating the process of the text recognition of the Arabic language in this indication. It is a figure explaining the text image containing an Arabic text. It is a figure explaining dividing a text image into a plurality of line images each including a plurality of pixels. It is a figure explaining the pixel and pixel value in a part of line image shown by FIG. 3A. It is a figure explaining the pixel and pixel value in a part of line image shown by FIG. 3A. It is a figure explaining the method of text feature extraction according to this application. FIG. 5 is a flowchart illustrating a text feature extraction process shown in FIG. 4. It is a figure explaining the other method of text feature extraction according to this application. It is a figure explaining other text feature extraction methods according to this indication. It is a figure explaining other text feature extraction methods according to this indication. It is a figure explaining other text feature extraction methods according to this indication. It is a figure explaining other text feature extraction methods according to this indication. FIG. 8 is a flowchart for explaining the text feature extraction process shown in FIGS. 7A to 7D.

DETAILED DESCRIPTION OF THE INVENTION FIG. 1 illustrates the general flow of Arabic text recognition according to the present invention. 1 to 3C, a text image 200 is acquired from an Arabic text document (step 110 in FIG. 1). The Arabic text in the text image 200 may be placed in a plurality of text lines 211-214, each of which includes a string of cursive Arabic characters. The text lines 211-214 are divided into a plurality of line images 311-313 (step 120 in FIG. 1). The line images 311, 312, or 313 are then divided into pixels 321-323, each assigned a pixel value (step 130 of FIG. 1). The width of the line images 311, 312, or 313 can be in the range between 2 and 100 pixels, or can be in the range between 3 and 10 pixels. Line images 311, 312, or 313 may include complete characters, partial characters, or combined characters.

  The pixel value represents the brightness value of the text image 200 at a specific pixel position. In some implementations, a high brightness value represents the color (or low density) of a bright image at pixels that may be located on a white background. A low brightness value represents the color (or high density) of a dark image that may be located within a stroke of an Arabic character. Pixel values may be expressed in different counting methods, such as binary, decimal, and hexadecimal.

  With reference to FIGS. 3A to 3C, the line image 311 includes an image portion 320 including a plurality of pixels 321 to 323. Each of the pixels 321 to 323 is assigned a binary pixel value “0” or “1”. Pixel value “1” represents a white background. The pixel value “0” represents a dark image color (that is, a low brightness value) in one stroke of Arabic characters. The disclosed system and method can also be adapted to multi-bit pixel values expressed in binary numbers, where the multi-bit pixel values expressed in binary numbers can have multiple tone levels (eg, (Grayscale) can represent image density.

  In accordance with this disclosure, text feature vectors may be extracted from text lines 211 or line images 311-313 (step 140 of FIG. 1). Details of various implementations of text feature extraction are discussed below in connection with FIGS. The exact form of the text feature vector can vary depending on the extraction method, as described below.

  The feature vector obtained in step 140 is then sent to a hidden Markov model (HMM) (step 150 in FIG. 1). In this disclosure, the HMM may be implemented by a hidden Markov model toolkit (HTK), which is a portable toolkit for building and manipulating hidden Markov models. HTK has no vocabulary and relies on models and grammar from sample characters for learning. The HMM provides a probability interpretation and can tolerate changes in the pattern found in the feature vector. Most of the functionality of HTK can be incorporated into library modules available in C source code. These modules are designed to operate with a conventional command line type interface, thus simplifying the writing of scripts for controlling the execution of the HTK tools.

  The HMM can be trained by using a feature vector acquired from a text image containing a known Arabic word (data transcription) (step 160 in FIG. 1). HTK is provided with a character model and ground truth for learning samples. A component for character modeling uses a feature vector and a corresponding ground truth to evaluate a character model. Observations generated by the training samples are used to adjust model parameters, while observations generated by the test samples are used to investigate system performance. Each state of the model represents a letter in the alphabet set, and each feature vector corresponds to one observation. The HTK learning tool can adjust the character model parameters using the prepared learning data and predict a known data transfer.

  HMM parameters were estimated from ground truth for the learning image segment. This segmentation can also be applied to contours to find segmentation points, extract features from these segments, and communicate feature vectors to the observation sequence. Segmentation-based techniques are used for dynamic programming to match word images and character strings. In the learning stage, a scanned line of text integrated with Grand Truth, which is text corresponding to a text image, is obtained as input. Each line is divided into narrow vertically divided windows from which feature vectors are extracted.

  The learned HMM is used to recognize the Arabic text in the feature vector using a dictionary and a language model (step 170 in FIG. 1). The recognition phase follows the same process of extracting feature vectors used with different knowledge sources estimated in the learning phase to find the highest likelihood character sequence. A recognition tool requires a network to describe the transition probabilities from one model to another. A dictionary and language model can be entered into the tool to help the recognizer output the correct state sequence.

  In some embodiments, referring to FIGS. 3A-5, line images 311-313 are digitized into an array of pixels 321-323, each characterized by a pixel value (step 510 of FIG. 5). ). The line image 311 is divided into a plurality of cells 410-460 as shown in FIG. 4 (step 520 in FIG. 5). Each of cells 410-460 includes a group of adjacent pixels, such as a 3 × 3 pixel array. For example, cell 420 includes pixels 422, 423 and other pixels.

  Next, the pixel value of each cell is represented by a binary cell number (step 530 in FIG. 5). The pixel value in each cell is first serialized. For example, the nine pixels 322-323 in cell 420 are serialized in the order of three consecutive rows as follows: 1,1,1,1,0,0,1,0,0. The series of binary pixel values is then mapped to a 9-bit binary cell number. The pixel value of pixel 322 is mapped to the most significant bit, and the pixel value of pixel 323 is mapped to the least significant bit. As a result, the pixel value in the cell 420 is represented by a 9-bit cell number 111100100 represented by a binary number. Similarly, the pixel values in cells 410-460 are converted to cell numbers 480 expressed in binary numbers, each in the range between 0 and 511.

  The binary cell number in the cell of the line image 311 is then converted to a decimal cell number 490 (step 540 in FIG. 5). The decimal cell number 490 is then serialized to form the feature vector for the line image 311 (step 550 in FIG. 5). Steps 520-550 are repeated for another line image. The feature vector from another line image 311-313 is then sent to the hidden Markov model to recognize Arabic characters in the text line (step 560 in FIG. 5).

  The above extraction method described in conjunction with FIGS. 4-5 represents an implementation of text feature extraction for the process described in FIG. It should be understood that the above text feature extraction method is compatible with multi-bit pixel values and other numerical representations in a data string. For example, pixel values can be represented by 3-bit or 5-bit binary numbers that can capture grayscale information (or multitones) in a text image. Multi-bit pixel values can improve the accuracy of the description of text features along the edges of the stroke.

  Furthermore, instead of binary numbers, pixel values can be represented by any numerical range between the minimum and maximum values. In some implementations, the pixel value can be a value that is proportional (or normalized) to a predetermined range, such as [0, 1] or [-1, 1]. The pixel value can then be quantized. A feature vector may be obtained as in steps 530-550.

  In some embodiments, referring to FIG. 6, the line image 610 is reduced in resolution (ie, miniaturized), thereby forming a miniaturized line image 620. For example, the line image 610 may have a height of 60 pixels. The miniaturized line image 620 may be 1/3 times as large and have a height of 20 pixels. The miniaturized line image 620 is digitized to form an array 630 of pixels, each represented by a pixel value. The pixel values for each column in array 630 are serialized to form a binary number. Binary numbers from different columns form a data string 640 that forms a feature vector. The feature vector obtained from the line image of the text line is sent to the hidden Markov model, whereby Arabic characters in the text line can be recognized (step 560 in FIG. 5).

  Referring to FIGS. 7A, 7B, and 8, the line image 700 is digitized into an array of pixels (step 810 in FIG. 8), similar to step 510 (FIG. 5). Pixels are arranged in a plurality of columns. The pixel value is represented by a single bit binary number having the value “1” or the value “0”. The pixel values in each column are serialized to form a single bit binary column (step 830 in FIG. 8).

  Next, as shown in FIGS. 7C and 7D, the frequency of consecutive pixels having the same binary pixel value of value “0” and value “1” is calculated (step 840 of FIG. 8). . The frequency is counted up to the cut off transition number. The frequencies are tabulated to form frequency counts 750 and 760 (step 850 in FIG. 8). Complementary pixel values, for example

In order to distinguish between two pixel columns that have the same number of transitions, a frequency count is performed by starting counting the number of values “1” from the topmost pixel of the column. In the left column, initially, the count of the pixel value “1” is “0”, followed by the pixel value “0” of “3” count. Complementary pixel values in the two columns result in the following frequency count:

It should be understood that the initial pixel count at the beginning of each column can also be performed for the pixel value “0” without departing from the spirit of the present invention.

  Each row in tabular frequency counts 750, 760 (FIGS. 7C, 7D) is from a white background (having pixel value “1”) to a dark text region (having pixel value “0”) or vice versa. Of the pixel value. To compress the data, the frequency count is truncated at the maximum transition number.

  The frequency counts in each column of tabular frequency counts 750, 760 form a feature vector (step 860 in FIG. 8). Therefore, in this embodiment, each column can also be referred to as a vector. Feature vectors from various columns in the line image are sent to the hidden Markov model (step 870 in FIG. 8).

  The maximum transition number is determined by statistical analysis on a large sample of Arabic text. As shown in Table 1, approximately 99.31% of the columns have 6 or fewer transitions. In other words, the majority of text images can be appropriately characterized by selecting 6 as the cut-off transition number.

  When building an HMM-based system, the type of feature vector used in learning and testing the system is first defined. Feature vectors can be classified into continuation types and separation types. In systems that use continuation type feature vectors, an array of coefficients sent to the model, and in some cases a matrix, is used. In systems where separate type feature vectors are utilized, a single coefficient is sent to the model. The vector quantization means converts the continuation type vector into the separation type vector, which is done by using the HQuant tool and the HCopy tool associated with HTK. HQuant is used to build a codebook from training data that will be used later with HCopy to generate separable type vectors. The construction of the code book affects the performance of the system depending on the size of the system, and is also affected by the amount of data used for the construction. HQuant uses a Linear Vector Quantization Algorithm to build a codebook, which is a computationally expensive algorithm for computation. In this disclosure, a new method named Unique Vector Quantization (UVQ) is introduced, which reduces computation time and improves system performance. This method reduces the number of feature vectors used to build a codebook that uses the Linear Vector Quantization Algorithm by eliminating feature vector iterations, and only for each feature vector. The focus is on reducing the number of feature vectors used to hold one copy. As shown in Table 2, the number of feature vectors in the corpus is greatly reduced.

  When trying to build a codebook using all of the 2000 different line image feature vectors, it was discovered that the maximum size that could be built for this codebook was 728. While it took 1 hour and 30 minutes to build a code book of 1024 size from only the unique feature vector, it took about 9 hours to build this code book. The recognition speed from these experiments using mono models is shown in Table 3. When a unique feature vector is used with a linear vector quantization algorithm, the size of the codebook increases. The calculation speed increased 6 times and the recognition speed increased.

  It should be understood that the method described above is not limited to the specific examples mentioned. The settings can be changed without departing from the spirit of the invention. For example, the foot cut transition number can be selected other than 6. The height and width of the line image can be different from the above example, as can the size of the cells in the line image. The form of the text feature vector can be changed according to the extraction method. For example, the feature vector may take the form of a number described in binary, decimal, or other notation.

  It should be thought that embodiment disclosed this time and its modification are illustrations in all the points, and are not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims. It is intended that the techniques disclosed in the embodiments and the modifications thereof can be implemented alone or in combination as much as possible.

Claims (16)

A computer-implemented method for automatically recognizing Arabic text,
Obtaining a text image containing lines of Arabic characters;
Digitizing the line of Arabic characters to form an array of two-dimensional pixels each associated with a pixel value represented in binary, the array of two-dimensional pixels comprising: A plurality of rows in the first direction and a plurality of columns in the second direction;
The method further comprises:
Counting the frequency of consecutive pixels of the same pixel value in a string of pixels in a column of pixels, each string of adjacent pixels having a different pixel value defined by a transition between them, Counting is further
When the number of transitions in a column reaches a predetermined cut-off transition number, stopping counting the frequency of consecutive pixels of the same pixel value in the column;
Forming a text feature vector using a frequency count obtained from a string in the column of pixels;
Recognizing a line of Arabic characters by sending the text feature vector to a hidden Markov model.   The computer-implemented method of claim 1, wherein the line of Arabic characters includes a plurality of Arabic words.   The computer-implemented method of claim 1, wherein the text feature vector is formed by a series of frequency counts obtained from a string of consecutive pixels in the column of pixels.   The computer-implemented method of claim 1, wherein the predetermined cut-off transition number is obtained by statistical analysis on Arabic text prior to the step of digitizing the line of Arabic characters. Method.   The computer-implemented method of claim 1, wherein the predetermined cut-off transition number is six.   The computer-implemented method of claim 1, wherein pixel values in the two-dimensional array are represented by a single bit binary number. Counting the frequency is
Assigning a value of a first frequency count to “0” when the pixel value of the first one or more pixels in the column is “0”, and after the first frequency count, The computer-implemented method of claim 6, wherein the column is followed by a number of consecutive pixels having a pixel value of “0”. Counting the frequency is
Assigning “0” as the value of the first frequency count when the pixel value of one or more pixels at the vertices of the column is “1”, and after the first frequency count, The computer-implemented method of claim 6, wherein the beginning of the column is followed by a number of consecutive pixels having a pixel value of “1”. A computer-readable program for causing a computer to execute the following, including a program code function, the program code function:
Get a text image containing a line of Arabic characters,
Digitizing the line of Arabic characters forms a two-dimensional array of pixels, each associated with a pixel value represented in binary, the two-dimensional pixel array having a first direction And a plurality of columns in the second direction,
The program code function further causes the computer to count the frequency of consecutive pixels of the same pixel value in a string of pixels in a column of pixels, and a string of adjacent pixels each having a different pixel value between them. The step of counting further stops counting the frequency of consecutive pixels of the same pixel value when the number of transitions in the column reaches a predetermined cut-off transition number. Including
The program code function is stored in the computer.
Forming a text feature vector using a frequency count obtained from a string in the pixel column;
Recognizing a line of Arabic characters by sending the text feature vector to a hidden Markov model.   The computer program according to claim 9, wherein the line of Arabic characters includes a plurality of Arabic words.   The computer program product of claim 9, wherein the text feature vector is formed by a series of frequency counts obtained from a string of consecutive pixels in the column of pixels.   The computer program product of claim 9, wherein the predetermined cut-off transition number is obtained by statistical analysis of Arabic text prior to the step of digitizing the line of Arabic characters.   The computer program according to claim 9, wherein the predetermined cut-off transition number is six.   The computer program according to claim 9, wherein pixel values in the two-dimensional array are represented by a single bit binary number. The step of counting the frequency comprises:
Assigning a value of a first frequency count to “0” when the pixel value of the first one or more pixels in the column is “0”, and after the first frequency count, The computer program according to claim 9, wherein the beginning of the column is followed by a number of consecutive pixels having a pixel value “0”. The step of counting the frequency comprises:
Assigning “0” as the value of the first frequency count when the pixel value of one or more pixels at the vertices of the column is “1”, and after the first frequency count, The computer program product of claim 9, wherein the number of consecutive pixels having a pixel value “1” follows at the beginning of the column.